[WS3] End-to-end logprob cross-benchmark tool

[Flink-ddd]

Part of #83 · Deferred to next phase — not in this month's sprint. Tracks #106.

The headline deliverable: one command runs the same fixed model + prompt set + seed through real vLLM and real Megatron, dumps both logprob streams, and computes drift. Tolerance reuses the #108 threshold table.

Planned PRs:

• One-command runner driving both real vLLM and real Megatron on the same fixed inputs.
• Dump + align both logprob streams; compute drift against the [WS1] Ground-truth harness + numerical contract for batch-invariant ops #108 thresholds.
• Drift report in the shared WS2 format ([WS2] Tolerance contract alignment & WS2 drift report format #116).